yingqing/core/backend/apps/ai/creation.py

"""全能创作 · 会话与消息的写入服务(契约 §1/§2)。

只放「怎么把一条消息安全落库」这类底座能力;agent 循环、工具执行、SSE 在
后续的 creation_agent.py 里,别混进来。
"""
from __future__ import annotations

from datetime import timedelta
import uuid

from django.core.cache import cache
from django.db import transaction
from django.db.models import Max
from django.utils import timezone

from .models import CreationConversation, CreationMessage


@transaction.atomic
def append_message(
    conversation: CreationConversation,
    *,
    role: str,
    kind: str = CreationMessage.Kind.TEXT,
    text: str = "",
    payload: dict | None = None,
    refs: list | None = None,
    task=None,
) -> CreationMessage:
    """往会话尾部追加一条消息,并刷新 last_active_at。

    seq 在事务里 select_for_update 锁住会话行再取 max+1 —— SSE 流式期间可能有
    并发写(用户抢答 / 轮询回填),不锁会撞 uniq_creation_message_seq。
    """
    locked = CreationConversation.objects.select_for_update().get(pk=conversation.pk)
    next_seq = (locked.messages.aggregate(m=Max("seq"))["m"] or 0) + 1
    message = CreationMessage.objects.create(
        conversation=locked,
        role=role,
        kind=kind,
        text=text,
        payload=payload or {},
        refs=refs or [],
        task=task,
        seq=next_seq,
    )
    locked.last_active_at = timezone.now()
    locked.save(update_fields=["last_active_at", "updated_at"])
    conversation.last_active_at = locked.last_active_at
    return message


def _assets_from_task(task) -> list[dict]:
    """任务落库的资产 → 结果卡要用的 {id,url,cover,type}。

    URL 必须走长期直链:预签名链 1 小时过期,写进消息 payload 第二天就打不开。
    """
    from apps.ai.services import asset_stable_url
    from apps.assets.models import Asset

    items = []
    assets = Asset.objects.filter(
        origin_task=task, is_deleted=False, purged_at__isnull=True,
    ).prefetch_related("files")
    for asset in assets:
        url, cover = asset_stable_url(asset)
        if not url:
            continue
        items.append({
            "id": str(asset.id),
            "url": url,
            "cover": cover or url,
            "type": "video" if asset.asset_type == Asset.Type.VIDEO else "image",
        })
    return items


def _meta_from_task(task, message: CreationMessage) -> dict:
    payload = message.payload or {}
    req = task.request_payload or {}
    model = ""
    if task.model_config_id:
        model = task.model_config.display_name or task.model_config.name
    return {
        "model": model or payload.get("model") or req.get("model") or "",
        "ratio": req.get("ratio") or payload.get("ratio") or "",
        "prompt": payload.get("prompt") or req.get("prompt") or "",
    }


def _message_task(message: CreationMessage):
    """GENERATING 消息挂的任务:优先 FK,payload.task_id 兜底(旧数据 / 序列化往返)。"""
    if message.task_id:
        return message.task
    task_id = (message.payload or {}).get("task_id")
    if not task_id:
        return None
    from .models import AITask

    return AITask.objects.select_related("model_config").filter(id=task_id).first()


@transaction.atomic
def fail_generating_message(message: CreationMessage, error: str) -> CreationMessage:
    """生成失败:GENERATING 原地改成 ERROR,不另开一条,避免中间态刷屏。"""
    is_person_reference = (message.payload or {}).get("kind") == "person_reference"
    message.kind = CreationMessage.Kind.ERROR
    message.text = (error or "生成失败")[:500]
    message.save(update_fields=["kind", "text", "updated_at"])
    conversation = message.conversation
    if is_person_reference:
        memory = dict(conversation.memory or {})
        memory["person_source_pending"] = False
        memory.pop("person_source_ready", None)
        conversation.memory = memory
        conversation.status = CreationConversation.Status.RUNNING
        conversation.agent_status = CreationConversation.AgentStatus.AWAITING_USER
        conversation.last_active_at = timezone.now()
        conversation.save(update_fields=[
            "memory", "status", "agent_status", "last_active_at", "updated_at",
        ])
        append_message(
            conversation,
            role="assistant",
            text="这次人物参考没生成成功。可以重新选择人物来源。",
            payload={
                "reply_hint": "选择一种方式继续…",
                "reply_options": [{"label": "重新选择人物", "text": "重新选择人物来源"}],
            },
        )
        return message
    conversation.status = CreationConversation.Status.FAILED
    conversation.last_active_at = timezone.now()
    conversation.save(update_fields=["status", "last_active_at", "updated_at"])
    return message


def sync_generating_message(message: CreationMessage) -> bool:
    """看挂着的 AITask 是否已经终态,是就把 GENERATING 改成 RESULT / ERROR。

    出图/出片在 worker 里跑,agent 只提交。前端轮询 GET 会话时靠这个回填;
    worker 结束时也会调一次,不用干等到下一次轮询。
    返回是否改了这条消息。
    """
    from .models import AITask

    if message.kind != CreationMessage.Kind.GENERATING:
        return False
    payload = message.payload or {}
    if payload.get("kind") == "video_segments":
        return _sync_segmented_video_message(message)
    task = _message_task(message)
    if task is None:
        return False
    if payload.get("kind") == "video_merge" and task.status in (AITask.Status.SUBMITTED, AITask.Status.POLLING):
        # Celery 不在线的本地环境同样可完成用户已确认的合并；此前从不在普通出片完成时调用。
        run_segmented_video_merge(str(task.id))
        task.refresh_from_db()
    if task.status == AITask.Status.SUCCEEDED:
        assets = _assets_from_task(task)
        if not assets:
            return False  # 状态已成功但资产还没落(极端竞态),下轮再试
        finish_generating_message(message, assets=assets, meta=_meta_from_task(task, message))
        return True
    if task.status in (AITask.Status.FAILED, AITask.Status.CANCELLED):
        from .generation_errors import public_error_for_task

        public_error = public_error_for_task(task)
        fail_generating_message(
            message,
            public_error.fallback_message if public_error else (task.error_message or "生成失败"),
        )
        return True
    return False


def _sync_segmented_video_message(message: CreationMessage) -> bool:
    """同步全能创作的多段出片。

    每段仍是普通 FREE_VIDEO 任务，故可沿用既有计费、轮询和资产落库；这里仅把它们聚合成
    一张结果卡。任一片段先完成就先写回 GENERATING 卡供用户预览；所有分段完成后才转成
    可合并的结果卡，绝不在此处触发 ffmpeg。
    """
    from django.conf import settings

    from .free_video import IN_FLIGHT_STATUSES, finalize_free_video, submit_free_video
    from .models import AITask

    # 团队级锁：同一团队若有两支长视频同时回填，不能都按同一份剩余并发额度补交。
    lock_key = f"omni:segment-schedule:{message.conversation.team_id}"
    if not cache.add(lock_key, "1", timeout=90):
        return False
    try:
        message.refresh_from_db()
        payload = dict(message.payload or {})
        segments = [dict(item) for item in payload.get("segments") or [] if isinstance(item, dict)]
        ids = [str(value) for value in payload.get("task_ids") or [] if value]
        if not ids or not segments:
            return False
        by_id = {
            str(task.id): task
            for task in AITask.objects.filter(id__in=ids).select_related("model_config")
        }
        tasks = [by_id.get(task_id) for task_id in ids]
        if any(task is None for task in tasks):
            fail_generating_message(message, "分段视频任务不完整，请重新生成。")
            return True

        # 本地没有 worker 时，会话轮询也能推进已提交片段；已在 worker 中则幂等返回。
        refreshed = []
        for task in tasks:
            assert task is not None
            if task.status in (AITask.Status.SUBMITTED, AITask.Status.POLLING):
                try:
                    task = finalize_free_video(task=task)
                except Exception:  # noqa: BLE001 - 单段网络抖动不能使整组直接失败
                    task.refresh_from_db()
            else:
                task.refresh_from_db()
            refreshed.append(task)

        segment_by_task = {
            str(item.get("task_id")): int(item.get("index") or index)
            for index, item in enumerate(segments, start=1)
            if item.get("task_id")
        }
        failed = next(
            (task for task in refreshed if task.status in (AITask.Status.FAILED, AITask.Status.CANCELLED)),
            None,
        )

        assets: list[dict] = []
        completed_segments = 0
        for task in refreshed:
            if task.status != AITask.Status.SUCCEEDED:
                continue
            task_assets = _assets_from_task(task)
            if not task_assets:
                continue
            segment_index = segment_by_task.get(str(task.id), completed_segments + 1)
            completed_segments += 1
            for asset in task_assets:
                assets.append({**asset, "label": f"第 {segment_index} 段", "segment_index": segment_index})
        assets.sort(key=lambda item: int(item.get("segment_index") or 0))

        if failed is not None:
            from .generation_errors import public_error_for_task

            number = segment_by_task.get(str(failed.id), 1)
            public_error = public_error_for_task(failed, operation="video_generate")
            detail = public_error.fallback_message if public_error else (failed.error_message or "请重试")
            error_text = f"第 {number} 段生成失败：{detail}"
            # 取消仍在飞的其它分段，避免失败后继续扣费/刷进度。
            for task in refreshed:
                if task.id == failed.id:
                    continue
                if task.status in IN_FLIGHT_STATUSES:
                    from apps.billing.services.ledger import release_credit

                    task.status = AITask.Status.CANCELLED
                    task.error_message = "同组其它分段已失败，已取消"
                    task.completed_at = timezone.now()
                    task.save(update_fields=["status", "error_message", "completed_at", "updated_at"])
                    if task.credit_reservation is not None:
                        try:
                            release_credit(
                                reservation=task.credit_reservation,
                                reason="同组分段失败，取消未完成片段",
                            )
                        except Exception:  # noqa: BLE001
                            pass
            if assets:
                first = next((task for task in refreshed if task.status == AITask.Status.SUCCEEDED), failed)
                finish_generating_message(
                    message,
                    assets=assets,
                    meta={
                        **_meta_from_task(first, message),
                        "kind": "video_segments",
                        "segment_count": len(segments),
                        "total_duration": payload.get("total_duration") or "",
                        "needs_merge": False,
                        "partial_failure": True,
                        "failed_segment_index": number,
                        "error": error_text,
                        "segments": segments,
                        "task_ids": ids,
                    },
                )
                conversation = message.conversation
                conversation.status = CreationConversation.Status.FAILED
                conversation.save(update_fields=["status", "updated_at"])
                append_message(
                    conversation,
                    role="assistant",
                    kind=CreationMessage.Kind.ERROR,
                    text=f"{error_text}。已保留成功片段供预览，未合并成片；可重新确认方案再生成。",
                )
            else:
                fail_generating_message(message, error_text)
            return True

        # 释放出来的并发槽位自动补交下一批。首批与后续批次统一从 generation_spec 派生，
        # 因而复用同一份人物/商品参考、完整脚本和 seed。
        pending = [item for item in segments if not item.get("task_id")]
        if pending:
            in_flight = AITask.objects.filter(
                team=message.conversation.team,
                task_type=AITask.Type.FREE_VIDEO,
                status__in=IN_FLIGHT_STATUSES,
            ).count()
            slots = max(0, int(getattr(settings, "FREE_VIDEO_MAX_CONCURRENT", 3)) - in_flight)
            spec = payload.get("generation_spec") if isinstance(payload.get("generation_spec"), dict) else {}
            if slots and not spec:
                fail_generating_message(message, "长视频续段参数不完整，请重新生成。")
                return True
            if slots:
                from .creation_agent import build_segment_video_submit

                total_duration = int(payload.get("total_duration") or 0)
                submitted_now = 0
                try:
                    for segment in pending[:slots]:
                        task = submit_free_video(
                            team=message.conversation.team,
                            user=message.conversation.created_by,
                            params=build_segment_video_submit(spec, segment, total_duration),
                        )
                        task_id = str(task.id)
                        segment["task_id"] = task_id
                        ids.append(task_id)
                        request_payload = dict(task.request_payload or {})
                        marker = dict(request_payload.get("omni_segment") or {})
                        marker.update({
                            "index": int(segment.get("index") or 0),
                            "start": int(segment.get("start") or 0),
                            "end": int(segment.get("end") or 0),
                            "total_duration": total_duration,
                            "message_id": str(message.id),
                        })
                        request_payload["omni_segment"] = marker
                        task.request_payload = request_payload
                        task.save(update_fields=["request_payload", "updated_at"])
                        submitted_now += 1
                except ValueError as exc:
                    if not submitted_now:
                        fail_generating_message(message, f"后续分段提交失败：{exc}")
                        return True
                    # 逐段提交期间并发槽位被别的请求占用时，保留本轮已提交任务；余下片段下轮再补。
                payload["task_ids"] = ids
                payload["segments"] = segments

        progress = {
            **payload,
            "assets": assets,
            "completed_segment_count": completed_segments,
            "submitted_segment_count": len(ids),
        }
        all_submitted = all(item.get("task_id") for item in segments)
        all_succeeded = all(task.status == AITask.Status.SUCCEEDED for task in refreshed)
        # 刚补交的新任务不在 refreshed 内，因此必须同时校验提交数，不能提前完成结果卡。
        if not all_submitted or len(refreshed) != len(segments) or not all_succeeded:
            if progress != payload:
                message.payload = progress
                message.save(update_fields=["payload", "updated_at"])
                return True
            return False

        # 状态已成功但资产尚未落库时继续等待，避免最终成片缺段。
        if completed_segments != len(segments):
            if progress != payload:
                message.payload = progress
                message.save(update_fields=["payload", "updated_at"])
                return True
            return False

        first = refreshed[0]
        finish_generating_message(
            message,
            assets=assets,
            meta={
                **_meta_from_task(first, message),
                "kind": "video_segments",
                "segment_count": len(segments),
                "total_duration": payload.get("total_duration") or "",
                "needs_merge": True,
                "auto_merge": True,
                "segments": segments,
                "task_ids": ids,
            },
        )
        message.refresh_from_db()
        # 全部分段成功后由平台自动合并，不再等用户点「合并成片」。
        try:
            merge_task, _generating = start_segmented_video_merge(
                conversation=message.conversation,
                message=message,
                user=message.conversation.created_by,
            )
            try:
                from .tasks import merge_omni_video_segments_task

                merge_omni_video_segments_task.apply_async(args=[str(merge_task.id)])
            except Exception:  # noqa: BLE001
                import logging

                logging.getLogger(__name__).warning(
                    "omni auto video merge enqueue failed for %s", merge_task.id, exc_info=True
                )
        except ValueError:
            # 已在合并中或状态不允许时忽略，避免重复入队。
            pass
        return True
    finally:
        cache.delete(lock_key)


def start_segmented_video_merge(*, conversation: CreationConversation, message: CreationMessage, user):
    """建立合并任务（全部分段成功后由平台自动调用；也可由旧接口手动触发）。合并执行仍只在 run_segmented_video_merge。"""
    from .models import AITask

    payload = dict(message.payload or {})
    if message.kind != CreationMessage.Kind.RESULT or not payload.get("needs_merge"):
        raise ValueError("这条结果不需要合并")
    state = str(payload.get("merge_state") or "")
    if state in {"queued", "processing", "completed"}:
        raise ValueError("这组片段已经在合并中或已合并")
    task_ids = [str(value) for value in payload.get("task_ids") or [] if value]
    source_tasks = list(AITask.objects.filter(id__in=task_ids).select_related("model_config"))
    if len(source_tasks) != len(task_ids):
        raise ValueError("分段视频不完整，无法合并")
    model_config = source_tasks[0].model_config
    merge_task = AITask.objects.create(
        team=conversation.team,
        created_by=user,
        project=None,
        task_type=AITask.Type.EXPORT,
        status=AITask.Status.SUBMITTED,
        model_config=model_config,
        idempotency_key=f"omni-video-merge:{conversation.id}:{message.id}:{uuid.uuid4()}",
        request_payload={
            "feature": "omni_video_merge",
            "source_message_id": str(message.id),
            "source_task_ids": task_ids,
            "prompt": payload.get("prompt") or "",
        },
    )
    payload["merge_state"] = "queued"
    payload["merge_task_id"] = str(merge_task.id)
    message.payload = payload
    message.save(update_fields=["payload", "updated_at"])
    generating = append_message(
        conversation,
        role="assistant",
        kind=CreationMessage.Kind.GENERATING,
        payload={"task_id": str(merge_task.id), "kind": "video_merge", "prompt": payload.get("prompt") or ""},
        task=merge_task,
    )
    return merge_task, generating


def run_segmented_video_merge(task_id: str) -> None:
    """后台执行用户已确认的合并。任务创建不等于执行；仅此函数调用 ffmpeg。"""
    from .free_video import _store_free_video_media
    from .models import AITask
    from .services import asset_stable_url
    from .video_replace import _concat_shot_media
    from apps.assets.models import Asset

    task = AITask.objects.select_related("team", "created_by").filter(id=task_id).first()
    if task is None or task.task_type != AITask.Type.EXPORT:
        return
    source_message_id = str((task.request_payload or {}).get("source_message_id") or "")
    source_message = CreationMessage.objects.filter(id=source_message_id).first()
    if source_message is None:
        return
    try:
        with transaction.atomic():
            locked = AITask.objects.select_for_update().get(id=task.id)
            if locked.status not in (AITask.Status.SUBMITTED, AITask.Status.POLLING):
                return
            locked.status = AITask.Status.POSTPROCESSING
            locked.save(update_fields=["status", "updated_at"])
        urls: list[str] = []
        for source_id in (locked.request_payload or {}).get("source_task_ids") or []:
            source_asset = (
                Asset.objects.filter(origin_task_id=source_id, is_deleted=False, purged_at__isnull=True)
                .prefetch_related("files")
                .first()
            )
            url, _cover = asset_stable_url(source_asset)
            if not url:
                raise ValueError("有片段尚未准备好，暂时不能合并")
            urls.append(url)
        merged_bytes = _concat_shot_media(urls)
        _store_free_video_media(task=locked, video_bytes=merged_bytes)
        with transaction.atomic():
            locked = AITask.objects.select_for_update().get(id=task.id)
            if locked.status != AITask.Status.POSTPROCESSING:
                return
            locked.status = AITask.Status.SUCCEEDED
            locked.completed_at = timezone.now()
            locked.save(update_fields=["status", "completed_at", "updated_at"])
            source = CreationMessage.objects.select_for_update().get(id=source_message.id)
            source_payload = dict(source.payload or {})
            source_payload["merge_state"] = "completed"
            source.payload = source_payload
            source.save(update_fields=["payload", "updated_at"])
    except Exception as exc:  # noqa: BLE001 - 合并失败不影响已完成的分段视频
        import logging

        logging.getLogger(__name__).exception("omni video merge failed for %s", task_id)
        with transaction.atomic():
            locked = AITask.objects.select_for_update().get(id=task.id)
            if locked.status == AITask.Status.POSTPROCESSING:
                locked.status = AITask.Status.FAILED
                locked.error_code = "MergeError"
                locked.error_message = str(exc)[:500]
                locked.completed_at = timezone.now()
                locked.save(update_fields=["status", "error_code", "error_message", "completed_at", "updated_at"])
            source = CreationMessage.objects.select_for_update().filter(id=source_message.id).first()
            if source is not None:
                source_payload = dict(source.payload or {})
                source_payload["merge_state"] = "failed"
                source.payload = source_payload
                source.save(update_fields=["payload", "updated_at"])
    finally:
        # 复用既有「生成中 → 结果/错误」转换，生成的合成片会成为独立结果卡。
        fresh = AITask.objects.filter(id=task_id).first()
        if fresh is not None:
            sync_generating_for_task(fresh)


def sync_generating_messages(conversation: CreationConversation) -> int:
    """把一条会话里所有已结束的 GENERATING 回填。返回改了几条。"""
    pending = list(
        conversation.messages.filter(kind=CreationMessage.Kind.GENERATING)
        .select_related("task", "task__model_config")
    )
    return sum(1 for message in pending if sync_generating_message(message))


def sync_generating_for_task(task) -> int:
    """worker / poll 终态后:只扫挂在这个任务上的 GENERATING。失败不能向外抛。"""
    try:
        from django.db.models import Q

        marker = (task.request_payload or {}).get("omni_segment") or {}
        aggregate_message_id = str(marker.get("message_id") or "") if isinstance(marker, dict) else ""
        lookup = Q(task=task)
        if aggregate_message_id:
            lookup |= Q(id=aggregate_message_id)
        pending = list(
            CreationMessage.objects.filter(
                lookup, kind=CreationMessage.Kind.GENERATING,
            ).select_related("conversation", "task", "task__model_config")
        )
        return sum(1 for message in pending if sync_generating_message(message))
    except Exception:  # noqa: BLE001 — 回填失败不能把已经成功的出图任务打成失败
        import logging

        logging.getLogger(__name__).warning(
            "omni create: sync generating for task %s failed", getattr(task, "id", "?"),
            exc_info=True,
        )
        return 0


@transaction.atomic
def finish_generating_message(message: CreationMessage, *, assets: list[dict], meta: dict) -> CreationMessage:
    """把「生成中」原地改成「结果」(契约 §2:不新增消息,避免中间态刷屏)。

    重生成是**新开一条** GENERATING → RESULT,所以对话流仍然是往下叠加;
    这里改的只是同一次生成自己的中间态。
    """
    original_payload = dict(message.payload or {})
    is_person_reference = original_payload.get("kind") == "person_reference"
    message.kind = CreationMessage.Kind.RESULT
    message.payload = {**original_payload, **meta, "assets": assets}
    message.save(update_fields=["kind", "payload", "updated_at"])
    conversation = message.conversation
    if is_person_reference:
        from apps.assets.models import Asset, Model

        asset_id = str((assets[0] if assets else {}).get("id") or "")
        asset = Asset.objects.filter(
            id=asset_id,
            team=conversation.team,
            is_deleted=False,
            purged_at__isnull=True,
        ).first()
        if asset is None:
            return fail_generating_message(message, "人物参考已生成，但未找到可锁定的图片资产")
        cast_index = int(original_payload.get("cast_index") or 1)
        cast_total = int(original_payload.get("cast_total") or 1)
        cast_model_name = str(original_payload.get("cast_model_name") or "").strip()
        is_pet = "宠物" in str(conversation.preset or "")
        if not cast_model_name:
            if is_pet:
                cast_model_name = "平台生成宠物角色"
            elif cast_total > 1:
                cast_model_name = f"平台生成出镜人物{cast_index}"
            else:
                cast_model_name = "平台生成出镜人物"
        model = Model.objects.filter(
            team=conversation.team,
            portrait_asset=asset,
            is_deleted=False,
            purged_at__isnull=True,
        ).first()
        if model is None:
            model = Model.objects.create(
                team=conversation.team,
                created_by=conversation.created_by,
                name=cast_model_name,
                source=Model.Source.AI,
                portrait_asset=asset,
                description=(
                    "由全能创作生成并锁定的宠物角色。"
                    if is_pet else
                    "由全能创作生成并锁定的视频出镜人物。"
                ),
                metadata={
                    "feature": "omni_create",
                    "conversation_id": str(conversation.id),
                    "cast_index": cast_index,
                    "cast_total": cast_total,
                },
            )
        elif model.name != cast_model_name and cast_total > 1:
            model.name = cast_model_name
            model.save(update_fields=["name", "updated_at"])
        pin_refs(conversation, [{
            "type": "model",
            "id": str(model.id),
            "name": model.name,
            "cover": (assets[0] if assets else {}).get("cover") or (assets[0] if assets else {}).get("url") or "",
        }])
        memory = dict(conversation.memory or {})
        memory["person_source"] = "platform_generate"
        model_ids = [
            str(item)
            for item in (memory.get("person_cast_model_ids") or [])
            if str(item).strip()
        ]
        model_id = str(model.id)
        if model_id not in model_ids:
            model_ids.append(model_id)
        memory["person_cast_model_ids"] = model_ids
        memory["person_model_id"] = model_id
        pending = int(memory.get("person_cast_pending") or cast_total or 1)
        pending = max(0, pending - 1)
        memory["person_cast_pending"] = pending
        memory["person_cast_total"] = int(memory.get("person_cast_total") or cast_total or 1)
        if pending > 0:
            # 还有其它角色定妆图在生成：继续等待，不打断用户确认
            memory["person_source_pending"] = True
            memory.pop("person_source_ready", None)
            memory.pop("person_confirm_pending", None)
            conversation.memory = memory
            conversation.status = CreationConversation.Status.RUNNING
            conversation.agent_status = CreationConversation.AgentStatus.IDLE
            conversation.last_active_at = timezone.now()
            conversation.save(update_fields=[
                "memory", "status", "agent_status", "last_active_at", "updated_at",
            ])
            return message

        memory["person_source_pending"] = False
        memory["person_source_ready"] = True
        memory["person_confirm_pending"] = True
        conversation.memory = memory
        conversation.status = CreationConversation.Status.RUNNING
        conversation.agent_status = CreationConversation.AgentStatus.AWAITING_USER
        conversation.last_active_at = timezone.now()
        conversation.save(update_fields=[
            "memory", "status", "agent_status", "last_active_at", "updated_at",
        ])
        total = int(memory.get("person_cast_total") or 1)
        if is_pet:
            confirm_text = "宠物角色已生成。你看这只宠物合适吗？是否使用这个宠物角色继续创作？"
            reply_hint = "回复「使用这个角色」继续，或说明想要的宠物品种与外观…"
            regen = "重新生成一个宠物角色"
            upload = "我上传宠物参考图"
            upload_label = "上传其他宠物"
        elif total > 1:
            confirm_text = (
                f"已生成 {total} 位出镜角色定妆图。长视频会按这些图分别锁脸，"
                "避免前后形象漂移。你看这组角色合适吗？是否使用这些角色继续创作？"
            )
            reply_hint = "回复「使用这个角色」继续，或说明想调整哪一位…"
            regen = "重新生成一个角色"
            upload = "我上传人物参考图"
            upload_label = "上传其他人物"
        else:
            confirm_text = "人物参考已生成。你看这位角色合适吗？是否使用这个角色继续创作？"
            reply_hint = "回复「使用这个角色」继续，或说明想调整的地方…"
            regen = "重新生成一个角色"
            upload = "我上传人物参考图"
            upload_label = "上传其他人物"
        append_message(
            conversation,
            role="assistant",
            text=confirm_text,
            payload={
                "reply_hint": reply_hint,
                "reply_options": [
                    {"label": "使用这个角色", "text": "使用这个角色继续创作"},
                    {"label": "重新生成一个", "text": regen},
                    {"label": upload_label, "text": upload},
                ],
            },
        )
        return message
    conversation.status = CreationConversation.Status.COMPLETED
    conversation.last_active_at = timezone.now()
    conversation.save(update_fields=["status", "last_active_at", "updated_at"])
    return message


def pin_refs(conversation: CreationConversation, refs: list[dict]) -> list[dict]:
    """把本轮引用的实体并进「实体锁定」。按 (type, id) 去重,保留首次出现的顺序 ——
    每轮无条件带上它们,这是多次生成之间锁脸/锁商品的唯一手段(契约 §5)。
    """
    merged = list(conversation.pinned_refs or [])
    seen = {(item.get("type"), str(item.get("id"))) for item in merged}
    changed = False
    for ref in refs or []:
        ref_type, ref_id = ref.get("type"), ref.get("id")
        if not ref_type or not ref_id:
            continue  # 缺 type/id 的 ref 解析不出实体,直接丢
        key = (ref_type, str(ref_id))
        if key in seen:
            continue
        merged.append(ref)
        seen.add(key)
        changed = True
    if changed:
        conversation.pinned_refs = merged
        conversation.save(update_fields=["pinned_refs", "updated_at"])
    return merged


# ---------------------------------------------------------------- Agent 在途锁(一账号同时只整理一份方案)

AGENT_LOCK_TTL_SECONDS = 20 * 60
AGENT_STALE_MINUTES = 20
AGENT_BUSY_DETAIL = "当前已有对话正在整理方案，请等待完成后再试"


def agent_lock_key(team_id) -> str:
    return f"omni:agent:{team_id}"


def release_agent_lock(team_id, conversation_id=None) -> None:
    """释放团队级 agent 锁。传 conversation_id 时只在锁仍指向该会话时删,避免误伤。"""
    key = agent_lock_key(team_id)
    if conversation_id is None:
        cache.delete(key)
        return
    if str(cache.get(key) or "") == str(conversation_id):
        cache.delete(key)


def acquire_agent_lock(team_id, conversation_id) -> bool:
    """cache.add 原子占坑。成功返回 True;已有占用返回 False。"""
    return bool(cache.add(agent_lock_key(team_id), str(conversation_id), timeout=AGENT_LOCK_TTL_SECONDS))


def cleanup_stale_agent_planning(team) -> int:
    """把超时仍卡在 planning 的会话清回 idle,并尝试释放对应 Redis 锁。"""
    from django.db.models import Q

    cutoff = timezone.now() - timedelta(minutes=AGENT_STALE_MINUTES)
    stale = list(
        CreationConversation.objects.filter(
            team=team,
            agent_status=CreationConversation.AgentStatus.PLANNING,
            is_deleted=False,
        ).filter(
            # started_at 为空也当过期(异常写入)
            Q(agent_started_at__lt=cutoff) | Q(agent_started_at__isnull=True)
        )[:20]
    )
    cleared = 0
    now = timezone.now()
    for conv in stale:
        updated = CreationConversation.objects.filter(
            pk=conv.pk,
            agent_status=CreationConversation.AgentStatus.PLANNING,
        ).update(
            agent_status=CreationConversation.AgentStatus.IDLE,
            agent_started_at=None,
            updated_at=now,
        )
        if updated:
            release_agent_lock(team.id, conv.id)
            cleared += 1
    return cleared


def find_planning_conversation(team, *, exclude_id=None):
    """返回团队里仍在 planning 的会话(已先清过期)。"""
    cleanup_stale_agent_planning(team)
    qs = CreationConversation.objects.filter(
        team=team,
        agent_status=CreationConversation.AgentStatus.PLANNING,
        is_deleted=False,
    )
    if exclude_id is not None:
        qs = qs.exclude(pk=exclude_id)
    return qs.order_by("-agent_started_at", "-last_active_at").first()


def team_agent_busy(team, *, exclude_id=None) -> bool:
    """DB planning 或 Redis 锁任一占用 → 忙。exclude_id 用于同一会话续跑自检。"""
    if find_planning_conversation(team, exclude_id=exclude_id) is not None:
        return True
    holder = cache.get(agent_lock_key(team.id))
    if not holder:
        return False
    if exclude_id is not None and str(holder) == str(exclude_id):
        return False
    # 锁还在但持有会话已不在 planning → 孤儿锁,清掉放行
    still = CreationConversation.objects.filter(
        id=holder,
        team=team,
        agent_status=CreationConversation.AgentStatus.PLANNING,
        is_deleted=False,
    ).exists()
    if not still:
        release_agent_lock(team.id, holder)
        return False
    return True


def begin_agent_planning(conversation: CreationConversation) -> bool:
    """占 Redis 锁并把会话标成 planning。失败(冲突)返回 False,调用方回 409。"""
    team_id = conversation.team_id
    if team_agent_busy(conversation.team, exclude_id=conversation.id):
        return False
    if not acquire_agent_lock(team_id, conversation.id):
        # 锁被别人占了;若锁指向自己(重入),允许续写状态
        holder = cache.get(agent_lock_key(team_id))
        if str(holder or "") != str(conversation.id):
            return False
    now = timezone.now()
    CreationConversation.objects.filter(pk=conversation.pk).update(
        agent_status=CreationConversation.AgentStatus.PLANNING,
        agent_started_at=now,
        last_active_at=now,
        updated_at=now,
    )
    conversation.agent_status = CreationConversation.AgentStatus.PLANNING
    conversation.agent_started_at = now
    conversation.last_active_at = now
    clear_agent_cancel(conversation.id)
    return True


def finish_agent_planning(
    conversation: CreationConversation,
    *,
    awaiting_user: bool = False,
) -> None:
    """Celery finally:写回 agent_status 并释放锁。"""
    status = (
        CreationConversation.AgentStatus.AWAITING_USER
        if awaiting_user
        else CreationConversation.AgentStatus.IDLE
    )
    now = timezone.now()
    CreationConversation.objects.filter(pk=conversation.pk).update(
        agent_status=status,
        agent_started_at=None,
        last_active_at=now,
        updated_at=now,
    )
    conversation.agent_status = status
    conversation.agent_started_at = None
    release_agent_lock(conversation.team_id, conversation.id)
    # 正常结束也清掉取消标记,避免下一轮误判
    clear_agent_cancel(conversation.id)


def agent_cancel_key(conversation_id) -> str:
    return f"omni:agent:cancel:{conversation_id}"


def clear_agent_cancel(conversation_id) -> None:
    cache.delete(agent_cancel_key(conversation_id))


def is_agent_cancel_requested(conversation_id) -> bool:
    """Celery tool round 之间读:用户点了终止则 True。"""
    return bool(cache.get(agent_cancel_key(conversation_id)))


def request_agent_cancel(conversation: CreationConversation) -> bool:
    """用户终止整理方案:打 Redis 取消标,立刻 idle + 释放团队锁。

    仅当当前仍是 planning 时成功。Celery 在下一轮工具间隙看到标后停跑,
    已落库消息保留。返回 True=已受理, False=当时不是 planning(调用方回 409)。
    """
    # 先打标,再收态 —— worker 在长 LLM 调用回来后也能看见
    cache.set(agent_cancel_key(conversation.id), "1", timeout=AGENT_LOCK_TTL_SECONDS)
    updated = CreationConversation.objects.filter(
        pk=conversation.pk,
        agent_status=CreationConversation.AgentStatus.PLANNING,
    ).update(
        agent_status=CreationConversation.AgentStatus.IDLE,
        agent_started_at=None,
        last_active_at=timezone.now(),
        updated_at=timezone.now(),
    )
    if not updated:
        # 已经不是 planning(竞态完成/别人清了);仍留标一会儿无害,清掉避免脏状态
        clear_agent_cancel(conversation.id)
        conversation.refresh_from_db(fields=["agent_status", "agent_started_at"])
        return False
    conversation.agent_status = CreationConversation.AgentStatus.IDLE
    conversation.agent_started_at = None
    release_agent_lock(conversation.team_id, conversation.id)
    return True